JusttYashh Challenge Inference
Try to disprove the pattern before accepting it.
Read schemas in ../soulcatcher/reference.md.
Worked example verdict: see challenge section in learned-skills/examples/pipeline-output/procedure-draft.md
Input
- ProcedureDraft from infer-procedure
- BehavioralTraces and PatternReport for counterevidence
Adversarial Checks
Run each check. Record pass/fail and reasoning.
| Check | Fail action |
|---|---|
| Appears in only one context subtype? | weaken confidence or reject |
| Explained by task type, not user habit? | reject |
| Only 2 occurrences? | reject — not a workflow |
| Core step skipped in multiple sessions? | mark optional vs core, or refine |
| Simpler explanation fits? | prefer simpler, refine |
| Counterexamples outweigh matches? | reject |
Minimum Evidence Rule
Two occurrences is not a workflow. Require 3+ relevant sessions with core sequence unless user overrides.
Output Format
# Challenge Verdict: {procedure_title}
**Verdict:** accept | refine | reject
## Checks
| Check | Result | Notes |
|-------|--------|-------|
| Cross-context evidence | pass | 5 distinct decision contexts |
| Task-type explanation | pass | Counterexample search not required by task |
| Minimum occurrences | pass | 12/15 sessions |
| Step optionality | pass | Core steps stable; 3 sessions skipped counterexamples |
| Simpler explanation | pass | No shorter sequence fits |
## Counterevidence
- decision-analysis-13: decided on pros/cons alone (deviation noted, not dominant)
- decision-analysis-14: committed before downside surfaced (single instance)
## Recommendation
**accept** — Pattern is cross-context, frequent, and not explained by task structure alone.
Adjust confidence to 87% after 2 counterexamples.
Verdict Actions
- accept → proceed to user approval, then compile-skill
- refine → return to infer-procedure with constraints (e.g. mark counterexample search as optional)
- reject → try next PatternReport or report no durable workflow found
Demo Bundle Note
When running on learned-skills/examples/sample-traces/, the expected verdict for adversarial-decision-analysis is accept with confidence 87%.
When demonstrating rejection, use a hypothetical 2-session pattern or reference sessions 13–14 as insufficient alone.